We recently fixed a bug in Chromium, but it looks like it affects Firefox as well and is implicit in the web-platform-tests suite (key_invalid.htm)

var store = db.createObjectStore('store', {keyPath: 'keypath'});
var values = ['string1', 'string2'];
var obj = { get keypath() { return values.shift(); } };

store.put(obj);
store.openCursor().onsuccess = function(e) {
  console.log('key is: ' + e.target.result.key);
  console.log('value has: ' + JSON.stringify(e.target.result.value.keypath));
};

In FF33 the key is 'string1' and the value has 'string2'. Having these out of sync seems like a violation of the intent of the spec, although the letter of the spec does imply tests before cloning.

In Chromium (circa 38) we clone, then evaluate the keypaths, so the result is 'string1' / 'string1'. 

This manifests in the key_invalid.htm test since it provides invalid key values such as function(){}. In Chromium we return DataCloneError since we clone before evaluating, whereas the test looks for DataError since it assumes evaluation before cloning.